Supplier price lists that go live in hours instead of days

Business context

The client is a building materials wholesaler with about 40 staff. It sells timber, fixings, and site supplies to builders through an online catalog and a small inside sales team. Margins are thin, so a quote is only safe when the catalog cost matches the supplier's latest price. The wholesaler asked us to automate supplier price updates because keeping the catalog in step had become a second job for one admin.

Suppliers issue new price lists most weeks. Some send a PDF, some send a spreadsheet, and each uses its own column order and product codes.

The problem

Keeping up was a manual race against the next email. A supplier update would arrive and the catalog might not reflect it for three days. Staff copied rows into the admin panel one at a time, checking every product code by hand.

Two failures followed. Quotes went out with old prices, which ate into margin. And when a price landed on the wrong row, the error reached a customer, who then had to be told the quote was wrong. Nobody could measure either problem, because the admin panel recorded the final value and not how it got there.

What we built

We built AI agent workflows that read each supplier file, match its products to catalog entries, and apply the changes. A person still approves anything unusual. The approach echoes an earlier project where we categorized receipts for a bookkeeping practice: confident automation applies, and a human queue catches the rest.

Reading the file

The first step pulls product code, description, unit, pack size, and price out of whatever the supplier sent. A PDF and a spreadsheet go through the same pipeline, so the supplier does not have to change how it works.

Matching products

Matching is the hard part, not reading. Supplier codes rarely match catalog codes exactly, and the same product can be described three ways. The agent compares the code, description, and pack size, then proposes a match with a confidence score. Exact matches apply on their own. Anything below the threshold goes to a person with both records side by side.

Applying changes

A price change above a set percentage is held for approval. Smaller changes reach the catalog with a full change log: supplier, date, old price, new price, and who approved it. A daily summary lists what changed overnight and what is waiting.

Rollout

• Week 1: Catalog and supplier file audit. We listed every supplier, file format, and update frequency.
• Week 2 to 3: Building extraction for the three most common layouts.
• Week 4 to 5: Matching against eight months of historical files while the buyer checked every proposed match.
• Week 6: Turning on automatic application for confident matches and the approval queue for the rest.

Before and after

Step	Before	After
Update arrives	Landed in one person's inbox	Picked up by the pipeline
Product match	Checked by hand, row by row	Suggested with a confidence score
Price change	Applied whenever someone got to it	Held above threshold, applied below
Record	Final value only	Full change log with old and new price

Results

Within the first month, the gap between a supplier email and a live catalog price fell from about three days to under six hours. The admin who had done the entry recovered roughly fourteen hours a week, which she now spends on supplier queries and stock.

About 88% of price lines matched and applied without a manual step. The rest reached the buyer as a short review list, and most of those were new products that needed a catalog entry anyway. Customer-reported pricing errors fell to zero in the eight weeks after launch, down from an average of three a month.

The approval threshold was the part we tuned most. Set too low, the buyer faced a long queue every morning. Set too high, and a real cost increase slipped through. We settled at 7% and left it easy to change.

Lessons learned

• Product matching, not reading the PDF, was the hard part. We budgeted two weeks for extraction and one for matching. It should have been the other way around, and the first historical run proved it.
• Suppliers changed formats without telling anyone. One moved a price column and broke an early rule. The confidence score and the human queue caught it, which is why we kept both.
• The change log did more for trust than the automation did. The buyer could see exactly what the agent did, so she stopped checking every entry.